Training Data — qwen3.8-9b-cyber-exploit-agent
This is the exact dataset the shipped model was trained on (QLoRA r16/a16, 3 epochs, best ckpt by eval loss).
- train_all_v2_shipped.jsonl — 395 samples: 280 CyberGym train-config tasks (8 blacklisted oss-fuzz IDs removed, Elfsong eval-200 never trained on) + 33 XRPL samples x3 (code-verified gates F1-F22/D/E/N, real issue texts, no maintainer comments in user turns) + 16 own labs/boundary samples.
- trackA.jsonl — Track A source samples (280).
- labs/ + evidence/ — 14 locally compiled and triggered labs (ASan logs, Python RCE markers). No invented crashes.
- scripts/ — full reproducible pipeline (dataset builders, SFT, merge, GGUF chain, eval gates).
- inference_system.txt — the training system prompt; use it at inference.
- train_ids.json / eval_ids.json — task id lists (train minus blacklist / eval holdout).
Dataset gate at build time: 0 blacklist ids, 0 user-turn leak markers, 0 schema violations, G1/G2/G6/G7/G8 verdicts pinned.
